Re: [boost] [string] Yet another Unicode string class

Subject: Re: [boost] [string] Yet another Unicode string class
From: Jeff Flinn (TriumphSprint2000_at_[hidden])
Date: 2011-02-15 07:35:01

Mathias Gaunard wrote:
> On 15/02/2011 00:12, Howard Hinnant wrote:
>> On Feb 14, 2011, at 5:55 PM, John Bytheway wrote:
>>
>>> On 14/02/11 19:08, Scott McMurray wrote:
>>>> On Mon, Feb 14, 2011 at 01:53, Mathias Gaunard
>>>> <mathias.gaunard_at_[hidden]>  wrote:
>>>>>
>>>>> SBO makes moving a string costly.
>>>>>
>>>>
>>>> How bug a buffer does SBO usually use?
>>>
>>> The libc++ string uses 23 bytes (on a 64-bit architecture) by squishing
>>> the length into a single byte for short strings.
>>>
>>> I don't see how that makes moves costly, though.  The move constructor
>>> for this string simply copies the bytes and zeroes out the source; it
>>> doesn't even need to branch based on whether it's a long or short
>>> string.  Perhaps the concern is that such techniques are not defined
>>> behaviour?  Or that the use of unions might confuse the optimizer?
>>
>> It isn't so much the size of the buffer that matters, but rather the 
>> total number of words that need to be copied or otherwise manipulated 
>> (e.g. zeroed).  The libc++ string is 3 words (on 32 and 64 bit) and 
>> the cost of the move is proportional to 3.  If the libc++ string 
>> needed to move (for example) 6 words, then its move constructor would 
>> be twice as expensive, though still cheap compared to a copy of a long 
>> string.
> 
> I didn't know it was just 3 words, that's still very cheap.
> Using memcpy doesn't break the strict aliasing guarantee, so that's legal.
> 
> It could be even cheaper if it could be copied using SIMD instructions, 
> but then that would put restrictions on the alignment of string, which 
> one might not want.

This caught my eye. A couple of years ago I looked at MMX/SIMD - tested 
an example off of the codeguru website which out of the box showed 
MMX/SIMD was faster. I unrolled the loop of the non-MMX/SIMD code using 
duff's device and saw no difference in performance with or without 
MMX/SIMD. I'm just wondering in your experience is there a real benefit 
of MMX/SIMD even discounting the alignment issue.

Thanks, Jeff